AAC壓縮域翻唱歌曲辨識系統

摘 要

隨著多媒體壓縮技術以及網際網路的蓬勃發展，使用者可透過網際網路下載或分享各種多媒體影音資料，然而，其影音內容卻可能是受版權保護而使用者在不知情的狀況下因此觸法。在本論文中對於AAC音訊壓縮格式提出翻唱歌曲辨識系統，其目標為快速檢索到資料庫內的原唱版本。從商業的角度來看，對於音樂的版權及管理是相當重要的，另一方面，對於使用者可以找到不同版本的歌曲既是有趣又實用。

在我們所提出的壓縮域翻唱歌曲辨識系統中，直接透過部分解碼得到音訊串流中的改良式餘弦轉換係數，將其重新定義到西方樂理上的十二平均率，在特徵擷取的步驟可以降低完全解碼的過程，並且利用音段切割降低特徵的時間維度以提升比對效率。因此，我們整體的系統可以省下很多運算複雜度，而且目前系統辨識準確率在Top-1已達到60%。在實際應用層面，對於檢索大量已編碼過的多媒體影音資料，我們的系統提供一個快速且準確地搜尋方法。

Cover Song Identification in AAC Compression Domain

ABSTRACT

With the rapid development of multimedia compression technology and Internet in recent years, users can easily download or share any kind of videos or music from networks. However, the downloaded contents may be copyrighted. In this work, we propose a system which can automatically identify cover version songs in AAC compression domain. Our goal is to fast retrieve the original version songs in a large coded database. From the commercial perspective, it is important to detect cover songs for musical copyrights’ management and licenses. Besides, it is interesting and useful to find out all versions for a particular song.

In our proposed system, the modified discrete cosine transform (MDCT) spectral coefficients are directly used to represent 12-dimensional chroma feature without a fully decoding process. In addition, we utilize segmentation to reduce time dimension in feature space for promoting the matching efficiency. Our overall system can save a lot of computation complexity and reach approximately 60% accuracy in Top-1. In practical applications, our system provides a good solution for retrieval system with large amount of coded multimedia files.